Close

Siyu (Sylvia) Dai

Download Resume

About Me

Sylvia's passion is to elevate robots from mechanical creations that follow pre-programmed trajectories to truly cognitive robots, and bring intelligent assistive robots into everyone’s home. Following this goal, she joined MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) and have researched on different aspects of robotics including risk-bounded motion planning, reinforcement learning, robotic manipulation, active learning and autonomous driving. Her Masters research focused on chance-constrained motion planning of robotic manipulation. This research drew techniques from a variety of fields, including trajectory optimization, sampling-based motion planning, probabilistic motion planning and optimal control. In her PhD research, more effort has been devoted to developing user-friendly intelligent agents that are capable of accomplishing useful manipulation tasks without intensive human supervision or reward engineering. Leveraging techniques from information theory and curriculum learning, she proposed a form of task-agnostic intrinsic motivation that allows reinforcement learning agents to learn manipulation tasks from scratch, and also developed a technique for efficiently utilizing expert demonstrations during reinforcement learning. Sylvia likes robotics, and she has gained lots of experience in robotic system integrations after working on many different robots, including Baxter from Rethink Robotics, WAM from Barrett, and HSR from Toyota. Currently, Sylvia works at Amazon Robotics AI as an Applied Scientist with a focus on computer vision and machine learning.

Education

Massachusetts Institute of Technology (MIT)

Doctor of Philosophy (PhD) in Mechanical Engineering

  • Major: Robotics
  • Minor: Machine Learning
  • Cumulative GPA: 5.0/5.0
  • Teaching Assistant for Dynamics and Control I, Physics II, Seminar in Undergraduate Research
  • Interests: Robot Manipulation, Reinforcement Learning, Imitation Learning, Curriculum Learning, Motion Planning
  • PhD Thesis: Learning to Make Decisions in Robotic Manipulation

Massachusetts Institute of Technology (MIT)

Master of Science in Mechanical Engineering

  • Cumulative GPA: 5.0/5.0
  • Research Assistant in Model-based Embedded and Robotic Systems (MERS) Group, MIT Computer Science and Artificial Intelligence Laboratory (CSAIL)
  • Master Thesis: Probabilistic Motion Planning and Optimization Incorporating Chance Constraints

Shanghai Jiao Tong University (SJTU)

Bachelor of Science in Naval Architecture and Ocean Engineering

  • Cumulative GPA: 3.9/4.0 (90.6/100); Overall Ranking: 1/73
  • Bachelor Thesis: Numerical Reconstruction and Mechanism Analysis on Vortex-Induced-Vibration of Steel Catenary Riser Caused by Platform Movement
  • Awarded 2016 Excellent Bachelor Thesis (Top 1%) of SJTU

Shanghai Jiao Tong University (SJTU)

Bachelor of Business Administration (BBA)

  • Cumulative GPA: 3.74/4.00 (88.4/100)
  • Bachelor Thesis: Study of Strategy for Precision Marketing based on the WeChat Platform
  • Awarded A+ thesis of 2016 SJTU BBA graduates

University of California, Berkeley

Summer Session

  • GPA: 3.82/4.00
  • Courses: Introduction to Oceans; Social, Political and Ethical Environment of Business

Experience

Robotics AI, Amazon.com

Applied Scientist II

  • Computer vision and machine learning for robotic manipulation. Focusing on 3D scene understanding and depth estimation.

Robotics AI, Amazon.com

Applied Scientist Intern, Supervisor: Aaron Parness, Sisir Karumanchi

  • Proposed a visual prediction approach to estimate the outcome of robot actions based on visual input.

Honda Research Institute USA

Research Intern, Supervisor: David Isele

  • Developed an intention-aware decision making and motion planning framework for autonomous vehicles in dense traffic with human drivers (manuscript to be submitted to Intelligent Vehicles Symposium (IV) 2022).

Mitsubishi Electric Research Laboratories

Research Intern, Supervisor: Yebin Wang

  • Developed a hierarchical motion planning approach that provides real-time motion plans for autonomous valet parking systems with incomplete map information.

General AI Lab, Horizon Robotics

Research Intern, Supervisor: Wei Xu

  • Proposed an empowerment-driven intrinsic exploration approach that allows reinforcement learning agents to learn manipulation skills with only sparse extrinsic rewards from the environment.

Publications and Patents

Selected Publications

Selected Patents

Scholarships and Main Awards

Scholarships and Fellowships

Main Awards

Selected Research Projects

Automatic Curricula via Expert Demonstrations (ACED)

September 2020 - June 2021

Advisor: Brian C. Williams, Computer Science and Artificial Intelligence Laboratory, MIT

Goal: To develop an imitation learning algorithm that utilizes demonstrations in an efficient manner and allows robotic manipulators to learn common tasks with as few as one demonstration

  • Proposed Automatic Curricula via Expert Demonstrations (ACED), a reinforcement learning framework that automatically extracts a sequence of curricula from expert demonstrations in order to accelerate the learning process of complicated robotic manipulation tasks
  • Integrated ACED with well-known imitation learning algorithms including behavior cloning and generative adversarial imitation learning, and evaluated the combination's performance on common robotic tasks
  • Analyzed the influence of number of demonstrations, number of curricula and policy initializations on the performance of ACED in robotic pick-and-place tasks and block stacking tasks
    • View Submission

    Autonomous Valet Parking incorporating Incomplete Map Information

    May 2020 - August 2020

    Advisor: Yebin Wang, Mitsubishi Electric Research Laboratories

    Goal: To design a motion planning approach for autonomous valet parking systems in environments with unknown obstacles that can achieve both long-horizon route planning from the parking lot entrance to the parking spot and short-horizon motion planning that properly parks the vehicle without collisions

  • Designed a hybrid architecture that integrates high-level route planning and low-level tree-based motion planning
  • Proposed a form of heuristics based on route planner solutions as well as a motion primitive trimming technique in order to improve the time- and memory-efficiency of the low-level parking planner
  • Developed a reactive planning system that allows for real-time trajectory repairing and replanning when encountering previously unseen obstacles
  • Analyzed the performance of the proposed autonomous valet parking algorithm on 19 tasks in two different parking lot maps and compared it with state-of-the-art parking planners
    • View Publication

    Reinforcement Learning for Robotic Manipulation Tasks with Sparse Rewards

    September 2019 - September 2020

    Advisor: Brian C. Williams, Computer Science and Artificial Intelligence Laboratory, MIT

    Goal: To develop a reinforcement learning approach that encourages robots to learn basic manipulation skills through intrinsic exploration, and then transfer the skills to more complex tasks in new environments

  • Implemented 3 different intrinsic exploration approaches and evaluated their performance on object-lifting and pick-and place tasks in two different manipulation environments
  • Developed an empowerment-based intrinsic motivation that maximizes the conditional mutual information (MI) between actions and states and compared he performance of 3 different MI estimation approaches
  • Combined the empowerment-based intrinsic motivation with diversity-driven rewards and enabled the robotic manipulator to learn a diverse set of skills
  • Proposed a learning from demonstration framework that combines intrinsic exploration with inverse reinforcement learning to accomplish long-horizon compound tasks
    • View Publication

    Improving Chance-Constrained Motion Planning using Machine Learning Methods

    February - August 2019

    Advisor: Brian C. Williams, Computer Science and Artificial Intelligence Laboratory, MIT

    Goal: To develop an offline learning scheme that can provide faster online reaction time and more accurate collision risk estimation for chance-constrained manipulator motion planning

  • Compared the performance of kernel-based regression, random forest and neural networks in terms of improving the accuracy and efficiency of collision risk estimation given a probability distribution of robot states
  • Demonstrated a significant improvement on planning speed using neural networks in 1000 randomly sampled simulation test cases
    • View Submission

    Fast-reactive Risk-aware Robotic Motion Planning and Execution System Design

    October 2017 - January 2019

    Advisor: Brian C. Williams, Computer Science and Artificial Intelligence Laboratory, MIT

    Goal: To develop a risk-aware robotic motion planning system that accounts for system process noises and observation noises, and can quickly provide safe plans for robots with complicated dynamics but work under uncertainty, for instance underwater vehicles and human support robots

  • Improved and tested an implementation of LQR-RRT* algorithm, and explored approaches of building probabilistic roadmaps accounting for complicated system dynamics
  • Implemented the Linear Quadratic Gaussian Motion Planning (LQG-MP) algorithm on the 7-DOF Baxter arm
  • Developed a quadrature-based collision risk estimation approach and a risk reallocation method to facilitate chance constraints satisfaction for high-dimensional robotic planning tasks
  • Conducted 1000 simulation tests and showed significant collision reduction compared to deterministic solutions
  • Designed a risk-aware motion planning and execution system that can iteratively improve plans during execution time by incorporating the Iterative Risk Allocation (IRA) algorithm
    • View Thesis

    Motion Planning System Integration with the Toyota HSR Robot

    November 2017 - May 2018

    Advisor: Brian C. Williams, Computer Science and Artificial Intelligence Laboratory, MIT

    Goal: To develop the software interface for the Toyota HSR robot so that the Chekov motion planning system our team developed can be applied

  • Generated HSR OpenRAVE model and redefined the kinematic chain by adding the base rotational joint as a DOF to the manipulator, in order to apply 6D analytical inverse kinematics calculations
  • Developed the ROS-OpenRAVE interface for HSR in order to synchronize the planning system (in OpenRAVE) and the execution system (in ROS)
  • Integrated our "pick and place" plan dispatching system with the Toyota HSR "mock-hardware" control and simulation system
  • Investigated the TrajOpt collision cost issue when the robot is grasping objects and modified the TrajOpt algorithm to disable collision for grasping fingers

    • Trajectory Optimization in Robot Manipulation Motion Planning

    September 2016 - September 2017

    Advisor: Brian C. Williams, Computer Science and Artificial Intelligence Laboratory, MIT

    Goal: To analyze the strengths and weaknesses of the TrajOpt algorithm in robot motion planning, and to improve its performance by providing initial trajectories through sparse roadmaps

  • Created 4 representative environments and randomly sampled 5000 pairs of kinematically feasible and collision-free start and target poses for each test environment
  • Evaluated the performance of 4 sampling-based planners and the TrajOpt planner in those test cases; showed that sampling-based planners are not fast-reactive and TrajOpt alone has low success rate
  • Compared and analylzed TrajOpt's performance under different costs and constraints
  • Combined TrajOpt with a sparse multi-query roadmap approach, and the performance of this combined planner shows superiority over current planner in terms of speed and success rate
    • View Publication

    Analysis on the Lateral Vibration of Steel Catenary Riser (SCR) Caused by Platform Movement

    December 2015 - June 2016

    Advisor: Shixiao Fu, State Key Laboratory of Ocean Engineering, SJTU

    Goal: To reconstruct the experimental response of SCR under platform heave motion through Finite Element Method (FEM), and explore the mechanism for large-amplitude Cross-Flow (CF) lateral movement of SCR

  • Established SCR model in both ABAQUS and Orcaflex and compared numerical reconstruction results with experiment data, which validated the accuracy of the numerical models and the reliability of the conclusions
  • Conducted preliminary SCR dynamic response analysis which leaves out the influence of vortex-induced forces, proving that the contribution of dynamic buckling near the touchdown point (TDP) to SCR lateral vibration is only about 1%
  • Calculated the vortex-induced forces through mode superposition method and inverse FEM; added them to the numerical model and reconstructed about 95% of the lateral vibration, which is very close to experimental data
  • Analyzed the influence of platform heave frequency, frequency and amplitude of vortex-induced forces, vertical acceleration and vertical velocity of TDP and SCR top tension on the numerical reconstruction results
    • View Thesis
    View More Projects

    Design of Force-Feedback Experimental Apparatus for Vortex Induced Vibration Analysis

    March 2015 - November 2015

    Advisor: Shixiao Fu, State Key Laboratory of Ocean Engineering, SJTU

    Goal: To implement a hybrid experimental equipment by combining force-feedback with on-line numerical simulation, which allows the modeling of complex structural dynamics, while fully accounting for fluid-structure interaction

  • Minimized the phase lag of real-time force-signal filtering process by analyzing influencing factors (including cut-off frequency, sampling frequency and filter order) of the Butterworth-filter-induced phase lag
  • Simulated whole force-feedback loop using Simulink to test errors caused by Newmark-β algorithm
  • Estimated forces acting on slides and test cylinder by deriving unique method based on the Morison Equation in order to choose appropriate power and model of the servo motor
  • Drafted contract with Chengdu Pucuike to manufacture this widely reconfigurable equipment, which can model the free-vibration with forced-vibration experiment

    • Analysis of Long Memory and Scaling Behavior for Bulk Carrier Freight Rate Volatility

    September 2014 - November 2015

    Advisor: Feier Chen, Ocean Research Center, SJTU

    Goal: To investigate long-range correlation of 10-year Baltic Dry Index (BDI) time series and to study influence of seasonal trends and load capacity on BDI long memory

  • Conducted structural break tests with the Iterated Cumulative Sum of Squares (ICSS) algorithm and explored the origin of structural breaks by analyzing social and economic environment
  • Initiated the application of Multifractal Detrended Fluctuation Analysis (MF-DFA) method in the study of freight rate volatility, noting that it accommodates nonstationary time series
  • Discovered scaling behavior during MF-DFA tests
  • Analyzed the influence of noise reduction on long memory detection and eliminated switching points with Vondrak filter
  • Provided investment and risk management advice to both long-term and short-term participants in the freight rate market, including ship operators, charters, commodity traders and hedge funds
    • View Publication

    End-to-End Dynamic Response Analysis of Marine Flexible Slender Bodies (MFSB)

    May 2014 - October 2015

    Advisor: Xuesong Xu, Fluid Mechanics Research Laboratory, SJTU

    Goal: To analyze dynamic response of MFSB lateral oscillation and to design the optimal upper-end movement maneuver for re-entry operations, marine cable laying and deep-water towing tasks

  • Established the Flexible Segment Model (FSM) method, which discretizes a slender body into finite segments governed by node moment equilibrium
  • Compared numerical and experimental dynamic response results to validate the accuracy of the FSM approach
  • Conducted numerical analysis of 1500m MFSB and investigated the influencing factors of lower-end dynamic response, including upper-end oscillation period and amplitude, as well as density and flexibility of the MFSB
  • Discovered minimum oscillation positions along the MFSB under different upper-end oscillation conditions
  • Awarded "Outstanding Research Project of Shanghai Jiao Tong University" (top 3%)
    • View Project Report

    Experiment on Wave Run-up Phenomenon via Particle-Image-Velocimetry (PIV) Approach

    September 2014 - April 2015

    Advisor: Xinliang Tian, State Key Laboratory of Ocean Engineering, SJTU

    Goal: To understand influence of waves on columns of semi-submersible platform by studying mechanism of wave run-up phenomenon and to optimize air-gap design of semi-submersible platform by predicting air-gap response

  • Calibrated 25 sets of waves and conducted experiments on cylinders of 5 different sizes, each placed horizontally, obliquely and vertically, to accommodate different column cross-section geometries and sea states
  • Measured wave elevations and forces acting on test cylinders with wave probes and six-dimensional force transducers
  • Calibrated the angle and focus length of CCD camera as well as the position of laser generator, in order to observe the velocity field through photos of tracer particles
  • Analyzed the flow field with image processing techniques based on auto-correlation and cross-correlation theories

    • Leadership and Extracurricular Experience

    Graduate Student Council Academic, Research and Career Committee, MIT

    Co-Chair

    • Initiated a subcommittee that works with MIT senior administration to improve advisor-advisee relationship
    • Host academic related events, including panels and workshops on academia and industry job hunting

    Ashdown House, MIT

    Chair of Ashdown House Executive Committee (AHEC)

    • Led housing-related initiatives in MIT, including changes to housing allocation policies and resident food source problem after the closure of the nearby supermarket
    • Hosted weekly AHEC meetings; interviewed and recruited Ashdown officers
    • Initiated a volunteer appreciation system to improve the community engagement in Ashdown

    Graduate Association of Mechanical Engineers (GAME), MIT

    President

    • Recruited GAME officers; organized officer transitions and officer social events
    • Represented GAME to attend department meetings and discussed issues including faculty-student intercation and searching for new deparment head
    • Delivered speeches during department orientation
    • Arranged GAME meetings with department faculty and expressed students’ concerns

    Ashdown House, MIT

    Floor Officer

    • In charge of Ashdown third floor resident-related issues
    • Organized monthly floor social event

    Graduate Association of Mechanical Engineers (GAME), MIT

    Social Event Officer

    • Hosted the sixth Mechanical Engineering department annual gala, including venue reservation, funding application, publicity, etc.
    • Organized GAME weekly "Muddy Monday" event
    • Held monthly joint social events with other departments in MIT

    Alumni Council, SJTU

    Director

    • In charge of updating alumni address book, organizing Homecoming Day and searching for SJTU alumni worldwide

    SJTU 2016 Bachelor Degree Conferring Ceremony

    Representative of Graduates

    • Delivered a speech on behalf of 2016 graduates to express the gratitude for SJTU and the confidence in our new future

    SJTU Student Congress

    Representative

    • Collected students' opinions on the academic system and campus life, and presented to the Student Congress

    Model United Nations (MUN), SJTU

    Conference Delegate

    • Organized "2013 MUN Conference of SJTU and East China Normal University" with 100+ participants

    Skills